|
Jochen Tuchbreiter wrote:
>
> Hi,
>
> do any of the "bells & whistels" packages for LVS (like Piranaha/Webmonkey
> etc.) support load-adapted weighting of realservers yet ? I may be wrong but
> IMHO it should be pretty easy to write a script that gets a webpage from
> each realserver (if you are building a load-balanced webserver-cluster) ,
> measures the time this takes and then sets the weights according to response
> time. This could also be useful for clusters of proxies, clusters of
> nameservers ....
Load balancing is a feature of piranha (there are multiple algorythms to choose
from), but it does not work the way you are suggesting. It does not
"time" how long a fetch takes. There are options to base the balancing on
number of connections, a manually specified "weight" (to favor larger systems
over smaller ones), and cpu load.
Based on load, piranha will distribute http connections to the more available
server.
> Has anything like this been implemented yet / are there good reasons against
> doing this ?
There are several reasons for not doing it that way, as it is very subjective
to the environment rather than server performance. Different web pages are
going to respond with different results, plus networking (router hops, etc.)
would also come into play in Direct Routing and Tunneling setups.
You'd have to have a dedicated, predictable web page that all real servers
were tested with in order to make any determination on how performance was
involved. This would also interrupt the real web service activities taking
place (the act of measurement affecting the thing you are measuring).
CPU load, plus the number of http connections, would probably provide a more
applicable metric even though it may be a less direct measurement.
--
Keith Barrett
Red Hat Inc. HA Team
kbarrett@xxxxxxxxxx
|
